Anchor Divergences proposes context-sensitive image comparisons without changing embeddings
The researchers say distributions over text anchors can change how images are compared while keeping their embeddings fixed.
TLDR
The researchers behind Anchor Divergences argue that semantic similarity depends on context, although it is often studied using a fixed measure of vector similarity. They say distributions over text anchors can change how images are compared without changing their embeddings.
Combined views
54
1 Source, first seen ago
